An agent is built to extract customer feedback sentiment. You want to show the LLM how to classify it as ‘Positive’, ‘Neutral’, or ‘Negative’. Which few-shot design is most helpful?